AI Powered Early Warning System to Improve Patient Safety

Transcription

1 AI Powered Early Warning System to Improve Patient Safety Session #231, March 8, 2018 Shelley Chang, MD, PhD and Vibin Roy, MD, MBA Parkland Center for Clinical Innovation (PCCI) 1

2 Conflict of Interest Shelley Chang, MD, PhD & Vibin Roy, MD, MBA Have no real or apparent conflicts of interest to report. 2

3 Agenda Background Early Model: Auto-EWS 1.0 Next Generation: Auto-EWS 2.0 Case Studies Summary 3

4 Learning Objectives Learning Objective 1: Summarize the scope and scale of preventable hospital deaths and unplanned transfers to the ICU Learning Objective 2: Distinguish between real-time automated EMR-based prediction models and rapid response team (RRT) protocols Learning Objective 3: Analyze process of developing, validating & implementing automated prediction models/early warning systems Learning Objective 4: Evaluate potential for automated prediction models to improve the early detection of at-risk hospital patients 4

5 Background 5

6 Cardiopulmonary Arrests Annually ~209,000 In-Hospital CPA s with high mortality rates 6 AHA Heart Disease and Stroke Statistics, 2016 Update

7 Early Detection is the Goal National Patient Safety Goal: The early detection of physiologic deterioration in order to reduce in-hospital mortality and prevent unplanned transfers to the intensive care unit (ICU) 7 Joint Commision, NPSG

8 Rapid Response Teams Team of providers is summoned to the bedside to immediately assess and treat the patient with the goal of preventing intensive care unit transfer, cardiac arrest, or death 8

9 Rapid Response Team Call Criteria at Parkland Health & Hospital System (PHHS) 9

10 Mixed Evidence on RRTs 10

11 Early Warning Systems a Possible Solution? 11

12 Early Warning Systems Inadequate Require monitoring & activation by overburdened staff Fail to systematically monitor all patients Demonstrate only modest accuracy identifying patients at risk for CPA or death 12

13 What else can be done to proactively identify patients at risk for clinical deterioration? 13

14 Complexity of the Issue Predicting Out-of-ICU adverse events through predictive modeling is incredibly complex 14

15 Early Automated Model vs MEWS Observed rates of out-of-icu resuscitation and death events stratified by quintiles of risk in the automated model BMC Med Inform Decis Mak Feb 27;13:28. 15

16 Clinical Deep Learning 16

17 Takes Action Predictive model generates a risk score EWS Alert fired on high risk patients RAT Nurse Assesses Patient Monitor with no action taken Automated Early Warning Systems Workflow 17

18 Early Model: Automated EWS 1.0 o The program began at PHHS in August 2014 with version 1.0 of the Automated EWS software o Over this time period 338 alerts generated to RAT o Paused program in August 2016 to allow development of version

19 Definitions EWS first = Automated EWS alert was first trigger and there were no prior RAT visits in last 24 hours A Medicine service patient is defined as having spent time on the medicine service, medicine subspecialty, PM&R, family medicine anytime during admission. 19

20 EWS First Alert with Corresponding Actions Auto-EWS 1.0 live mode identified approximately 1 critically deteriorating patient on the medical service every 2 weeks that was not otherwise identified by staff. 57 cases first identified by Auto-EWS alert 20

21 Out-of-ICU CPA (Auto-EWS 1.0 Live Mode) Preliminary analysis shows we will need an additional 22 months of follow-up to have power to detect a 25% reduction in the hospital-wide out-of-icu CPA rate from 3.03 to

22 EWS 2.0: Feature Development and Testing 22

23 Auto-EWS 2.0 New Features o Auto-EWS 2.0 was prospectively tested in silent mode for 1-year prior to live mode deployment in July 2017 o Critical areas of improvement: 1. Improve alert sensitivity and specificity 2. Deliver the contextual reasons for the alert 3. Method to snooze alerts that don t warrant immediate actions 4. Further refined filter criteria 23

24 Improved sensitivity to detect adverse events in next 72-hours (Silent Mode Analysis*) Auto-EWS has ability to provide early warning* for: 63% of deaths 27% of code events 25% of unplanned ICU transfers 24 *Based on review of 337,833 6-hr at-risk periods from 20,813 inpatient encounters over a 1 year period of concurrent 1.0 (live) and 2.0 (silent mode) scoring

25 Clinical Contextual Reasons Provide the top 1-2 laboratory, vital sign, or organ system dysfunction drivers responsible for crossing fire/alert threshold Clinical contextual reasons are grouped into categories: ABG (related to PCO2, PO2, ph) Lung (related to Resp rate or SpO2) BP (Blood pressure related) Pulse (Pulse related) Trop (Troponin related) Infection (related to Lactate, Platelets, WBC, Temperature) LOC (related to LOC = level of consciousness) Blood (related to Hemoglobin, Hematocrit, Plats, INR) Liver (related to INR, Total bilirubin, Albumin) Kidney (related to BUN, Creat, CO2, Sodium) 25

26 Method to Snooze Alerts EWS RAT Nurse Assesses Patient via EHR RAT in-person Assessment Alert fired on high risk patients Fill out QI field with reason for not seeing patient 26 Monitor with no immediate action taken QI field filled = Snooze for X hours

27 Filter Criteria EWS Exclude Certain Patients Exclude Certain Locations Last RAT Call or Pieces Alert DNR & Comfort Care ICU, PCU, ER, Rehab, Psych, Radiology & Procedure Areas, Cath Lab, OR, L&D, Dialysis Hours since RAT alert. Hours since Pieces alert. Customized filter criteria to improve operational efficiency and RAT staff time PiecesQI Reason for Not Seeing Patient Time Since Location Bed & Room Assigned Hours since PiecesQI Reason recorded for Not Seeing Patient Hours since patient has been out of ICU or procedure suite before alert. Patient must have a bed/room assigned Team Will only generate alerts on Medicine patients 27

28 Potential Cost Savings Opportunities (Silent Mode Analysis*) Over a 12-month EWS 2.0 silent mode evaluation period, there were 234 patients identified early on by Automated EWS (but not by nurse call) who subsequently had an event (emergent transfer to ICU, CPA/ARC/Code, or Death) in the next 72 hours. * ICU costs estimated at $2,759 per day based on Becker Hospital Review ( 28

29 EWS 2.0 Live Mode: Early Results 29

30 EWS Identifies Critically Ill Patients (Auto-EWS 2.0 Live) A subset of patients first identified by EWS later required subsequent RAT activation by nurses These patients have very high subsequent ICU transfer rates and in-hospital mortality 30

31 More Time to Intervene Before It s Too Late (Auto-EWS 2.0 Live Mode) Most cases detected by Nursing Staff were followed-by immediate deterioration within 12 hours May be too late for some In contrast, Automated EWS more often provided advanced warning Earlier Initiation of Interventions 31

32 More Alerts Resulting in RAT Interventions* (Auto-EWS 2.0 vs 1.0 Live Mode) * Considers both immediate and subsequent actions taken in same admission ** Includes only cases with no recent nurse call in past 24 hours 32

33 Case Studies 33

34 Case # Auto EWS Activation (Triggers: BP & Infection) Pt noted to be hypotensive; had melena previously and received IVF bolus for hypotension; ICU consult was placed, resident was hesitant to transfer to ICU but agreed to do so after education from RRT RN. Shortly after arriving in ICU, pt became more hypotensive, had large bloody stool and large drop in blood counts (H/H 5.9/17.7). Pt was started on massive transfusion protocol, emergently intubated and taken to IR for mesenteric angiogram. Photo from: 34

35 Case #1 Lessons: Ability to identify patient who needed higher level of care Education of providers needed and RRT members are strong advocates Photo from: 35

36 Case # Auto EWS Alert (Trigger: BP) Patient BP had trended up to as high as 238/115. Patient was agitated and had history of dementia and aggressive behaviour. RAT advocated on behalf of the patient, administered Haldol and Hydralazine. BP came down to 144/66. 36

37 Case #2...worked great for this patient!...caught the BP and we were able to treat the pt s BP and hopefully prevented a stroke - Heather Wolf, RAT Coordinator 37

38 Summary Background: Challenging to predict which patients are at risk for adverse in-hospital events such as CPA s and ICU transfers Value of EWS: A real-time predictive model for identifying clinically deteriorating patients in the hospital and may help optimize mobilization of resources act prior to adverse events (CPA, respiratory failure, death) Silent mode Auto-EWS 2.0 showed significant improvement over the 1.0 version. Live Mode Auto-EWS 2.0 triggered RAT assessments of not otherwise detected patients which resulted in ~1 patient per week needing immediate transfers to ICU and ~1 patient every 2 days needing significant actions taken by RAT team 38

39 Acknowledgements Parkland Rapid Assessment Team Heather Wolf, Mitchell Netterville, Kelly Heathman & the RAT Team Lisa Mack & the Resuscitation Committee Drs. Carlos Girod and Matthew Leveno PCCI - Parkland Center for Clinical Innovation Ramon Depaz, Paula Olson, & Adeola Jaiyeola Monal Shah, Brian Lucena, & Chris Clark Parkland IT Dr. Brett Moran & Matthew Kull & John McManus Jacqueline Roberson & Raychel Lee Koushik Kaku (Kash) & Maurice King 39 Pieces Technologies Yi Cai & Kyle Morris & Alex Ramirez Greg Luptowski & Praseetha Cherian Drs. Anand Shah & Ruben Amarasingham

40 Questions For more information please contact Vibin Roy at or Shelley Chang at 40